KineBridge: Bridging sparse constraints and foundation motion priors
Authors: Xi, Y., Wang, X., Ye, X., Zhang, J.J., Yang, X.
Journal: Computers and Graphics
Publication Date: 01/12/2026
Volume: 141
ISSN: 0097-8493
DOI: 10.1016/j.cag.2026.104813
Abstract:Sparse pose keyframes together with a full root trajectory provide direct geometric control for human motion authoring, but the transitions between pose constraints remain inherently ambiguous. Task-specific generators must learn both constraint alignment and a broad motion distribution from limited control data, which often leads to over-smoothed transitions or local artifacts. We present KineBridge, a geometric-control framework built on the pretrained HY-Motion 1.0 generator. The foundation backbone remains frozen, while a parallel trainable branch converts masked pose and trajectory conditions into a temporally aligned control sequence. The branch separately encodes the two modalities, applies Sparse-KV, Dense-Query attention according to their validity masks, and uses bidirectional fusion to couple articulated pose with root-trajectory context. Zero-initialized residual blocks then inject the condition tokens throughout the frozen double-stream and single-stream Transformer stages. The original text inputs are replaced with the pretrained model’s fixed null text features, enabling the model to focus on geometric control while retaining the pretrained motion prior. On the HumanML3D/CombatMotion test split, KineBridge achieves the best MPJPE, Anchor MPJPE, FID, and diversity among the compared systems, while PMG remains stronger in trajectory consistency. On the cross-dataset AIST++ evaluation, it achieves the best keyframe-alignment, bone-length, foot-sliding, ground-penetration, and diversity results. Component and injection-frequency ablations further validate sparse-aware propagation, pose–trajectory fusion, and layer-wise residual conditioning. These results demonstrate that precise sparse-pose control can be integrated into a frozen foundation motion model without retraining the backbone.
Source: Scopus